COMBO

COMBO is jointly trained neural tagger, lemmatizer and dependency parser implemented in python 3 using Keras framework. It took part in 2018 CoNLL Universal Dependency shared task and ranked 3rd/4th in the official evaluation.

Paper

The COMBO description can be found here: Semi-Supervised Neural System for Tagging, Parsing and Lematization.

Usage

Training your own model:

python main.py --mode autotrain --train train_data.conllu --valid valid_data.conllu --embed external_embedding.txt --model model_name.pkl --force_trees

Making predictions:

python main.py --mode predict --test test_data.conllu --pred output_path.conllu --model model_name.pkl

Trained models

Models trained on UD dataset:

Language	Treebank	LAS	MLAS	BLEX	Model
Afrikaans	af_afribooms	84.72	72.91	74.98	377 MB
Ancient Greek	grc_perseus	74.20	53.30	54.29	101 MB
Ancient Greek	grc_proiel	76.45	59.95	67.47	101 MB
Arabic	ar_padt	71.95	62.75	64.38	737 MB
Armenian	hy_armtdp	28.15	5.02	11.25	738 MB
Basque	eu_bdt	83.12	68.82	77.96	737 MB
Bulgarian	bg_btb	89.36	81.10	79.98	738 MB
Buryat	bxr_bdt	15.16	1.09	1.92	90 MB
Catalan	ca_ancora	90.54	83.11	85.20	737 MB
Chinese	zh_gsd	63.92	53.48	57.84	744 MB
Croatian	hr_set	86.32	71.12	79.74	737 MB
Czech	cs_cac	90.72	83.27	86.69	740 MB
Czech	cs_fictree	91.83	84.23	87.81	740 MB
Czech	cs_pdt	90.34	84.04	86.96	740 MB
Danish	da_ddt	83.43	74.22	77.58	737 MB
Dutch	nl_alpino	87.15	74.93	77.06	737 MB
Dutch	nl_lassysmall	84.27	72.65	75.44	737 MB
English	en_ewt	82.31	73.33	76.52	737 MB
English	en_gum	82.82	73.24	73.57	737 MB
English	en_lines	80.33	72.25	74.01	737 MB
Estonian	et_edt	83.46	75.79	72.07	738 MB
Finnish	fi_ftb	86.89	78.42	81.06	739 MB
Finnish	fi_tdt	85.93	78.65	72.39	739 MB
French	fr_gsd	85.42	77.08	79.72	738 MB
French	fr_sequoia	88.99	81.48	84.67	738 MB
French	fr_spoken	74.31	63.43	65.34	738 MB
Galician	gl_ctg	81.17	68.15	73.60	736 MB
Galician	gl_treegal	73.21	52.88	62.86	736 MB
German	de_gsd	77.43	54.28	68.59	738 MB
Gothic	got_proiel	65.87	50.81	59.30	48 MB
Greek	el_gdt	88.49	76.15	78.57	738 MB
Hebrew	he_htb	63.69	50.26	53.58	737 MB
Hindi	hi_hdtb	91.43	76.23	86.29	593 MB
Hungarian	hu_szeged	79.47	66.09	72.51	737 MB
Indonesian	id_gsd	78.40	67.30	75.10	737 MB
Irish	ga_idt	69.24	37.31	47.32	206 MB
Italian	it_isdt	91.03	83.18	84.76	737 MB
Italian	it_postwita	73.99	61.14	62.98	737 MB
Japanese	ja_gsd	73.69	57.82	60.62	743 MB
Kazakh	kk_ktb	22.38	4.40	7.86	738 MB
Korean	ko_gsd	80.66	74.49	66.13	741 MB
Korean	ko_kaist	84.88	76.92	72.40	743 MB
Kurmanji	kmr_mg	21.95	2.26	05.01	45 MB
Latin	la_ittb	85.54	79.84	83.51	526 MB
Latin	la_perseus	68.07	49.77	52.75	526 MB
Latin	la_proiel	70.08	56.82	64.94	526 MB
Latvian	lv_lvtb	80.71	66.22	71.80	637 MB
North Sámi	sme_giella	57.16	39.66	45.03	47 MB
Norwegian	no_bokmaal	89.33	79.51	84.68	737 MB
Norwegian	no_nynorsk	88.36	79.32	82.89	737 MB
Norwegian	no_nynorsklia	68.26	57.51	60.98	737 MB
Old Church Slavonic	cu_proiel	71.14	56.52	66.04	48 MB
Old French	fro_srcmf	84.81	76.75	81.20	52 MB
Persian	fa_seraji	86.14	80.30	76.29	737 MB
Polish	pl_lfg	94.62	86.44	89.31	737 MB
Polish	pl_sz	91.38	80.45	85.59	737 MB
Polish	poleval2018	86.11	76.18	79.86	115 MB
Portuguese	pt_bosque	87.57	74.31	80.31	737 MB
Romanian	ro_rrt	85.31	76.84	79.54	737 MB
Russian	ru_syntagrus	91.10	85.37	87.16	741 MB
Russian	ru_taiga	74.24	61.59	64.36	741 MB
Serbian	sr_set	87.27	73.79	79.92	738 MB
Slovak	sk_snk	83.76	63.97	75.34	54 MB
Slovenian	sl_ssj	85.72	75.07	81.11	737 MB
Slovenian	sl_sst	58.12	45.93	50.94	737 MB
Spanish	es_ancora	89.68	82.60	84.51	737 MB
Swedish	sv_lines	81.97	66.26	77.01	737 MB
Swedish	sv_talbanken	85.89	77.68	80.74	737 MB
Turkish	tr_imst	63.54	52.51	58.89	737 MB
Ukrainian	uk_iu	84.71	69.88	77.97	738 MB
Upper Sorbian	hsb_ufal	21.30	1.45	4.53	139 MB
Urdu	ur_udtb	81.53	55.70	72.49	485 MB
Uyghur	ug_udt	63.10	40.71	52.76	165 MB
Vietnamese	vi_vtb	42.53	35.11	38.47	736 MB

License

CC BY-NC-SA 4.0

Citation

@InProceedings{rybak-wrblewska:2018:K18-2,
  author    = {Rybak, Piotr  and  Wr{\'{o}}blewska, Alina},
  title     = {Semi-Supervised Neural System for Tagging, Parsing and Lematization},
  booktitle = {Proceedings of the {CoNLL} 2018 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies},
  month     = {October},
  year      = {2018},
  address   = {Brussels, Belgium},
  publisher = {Association for Computational Linguistics},
  pages     = {45--54},
  url       = {http://www.aclweb.org/anthology/K18-2004}
}